Proposes CodeQuant for quantizing mixture-of-experts models by combining clustering and quantization to handle outlier-induced errors.
Presents IceCache for memory-efficient KV-cache management in long-sequence LLMs via CPU offloading and selective GPU retention.
Proposes WaveMoE, a mixture-of-experts foundation model for time series forecasting using wavelet-enhanced frequency-domain information.
Studies how reward hacking during RLHF fine-tuning degrades LLM calibration and uncertainty quantification despite improving helpfulness.
Explores online continual self-supervised learning with focus on stability-plasticity trade-off in models learning from unlabeled streaming data.
MoEITS: green AI approach for reducing computational burden of Mixture-of-Experts LLMs through simplification.
Machine unlearning method for removing training data influence without direct access to forget sets.
Spectral analysis of LoRA weight updates showing low-frequency dominance enables efficient parameter-efficient fine-tuning.
Federated learning framework for IoT networks with energy efficiency optimization for small-scale datasets.
Self-distillation method for multi-turn LLM agents using skill-conditioning to improve sample efficiency in reinforcement learning.
On-policy distillation method for LLM alignment with adaptive weighting based on signal quality and credit assignment.
Communication-efficient optimization method extending Muon for federated learning of large language models.
Revisits value modeling in LLM reinforcement learning using generative critics for improved credit assignment.
Transformer architecture that dynamically determines its own depth and width during training by pruning redundant heads.
Reinforcement learning benchmark for Pokemon Red game with long horizons, sparse rewards, and complex control mechanics.
Improved online covariance estimation for averaged SGD with minimax-optimal convergence rates via trajectory regression.
Theoretical framework explaining how transformers learn in-context via mirror descent over mixture of transition distributions.
Proposes readiness indices based on Task2Vec embeddings to predict federated learning performance before training.
Establishes first information-theoretic lower bounds for score query complexity in diffusion model sampling.
Graph neural network domain adaptation method using information bottleneck and online distillation for robustness to distribution shifts.
Theoretical analysis of in-context learning in transformers beyond stationary settings, explaining how models adapt without parameter updates.
Subset selection framework using optimal transport for prototype selection with better handling of minority classes.
Novel framework for hypergraph neural networks using PDE-inspired diffusion equations to address oversmoothing and improve message passing.
Theoretical analysis of continuous-time online learning with two-layer neural networks in diffusion environments, establishing regret bounds.
Scheduling algorithm for LLM inference with provable stability when decode lengths are unknown, addressing memory overflow challenges in production systems.
Theoretical result showing K-way energy probes in predictive coding networks reduce to softmax, explaining apparent richness of per-hypothesis energy signals.
Rollout tree-based credit assignment method for multi-step agentic RL, leveraging implicit state overlap between group rollouts to avoid uniform advantage assignment.
Analysis of credit assignment in reinforcement learning with verifiable rewards using polarity-entropy decomposition to diagnose token update patterns in LLM reasoning.
Benchmark evaluating mechanistic interpretability methods under conditions where model explanations are absent, controlling for elicitation confounding effects.
Optimization technique for continual learning reducing computational overhead of C-Flat while maintaining ability to balance new and old task performance.
Method for detecting LLM hallucinations using counterfactual graph intervention to identify causal mechanisms, moving beyond passive signal-based classification approaches.
Bottleneck tokens framework for unified multimodal retrieval in decoder-only MLLMs, providing explicit pooling and token-level guidance for embedding alignment.
Class-incremental learning approach using quantum-gated knowledge distillation to address catastrophic forgetting in pretrained models across streaming task sequences.
Distributionally robust variant of k-means clustering using Wasserstein-2 balls to protect against outliers, distribution shifts, and limited sample sizes.
Meta's approach to reducing LLM hallucination in enterprise workflows by framing mitigation as Minimum Bayes Risk problem, critical for legal and compliance applications.
Compression pipeline for federated learning integrating pruning, quantization, and coding techniques to reduce communication and computational overhead in constrained environments.
Parameter-free algorithms for unconstrained online learning with regret bounds scaling with gradient variation, requiring no prior knowledge of model parameters.
Unified graph prompt learning framework using low-rank message prompting for efficiently fine-tuning pre-trained graph neural networks across multiple graph components.
Streaming approximate nearest neighbor index for high-dimensional vectors inspired by mycelium growth patterns, using adaptive topology with edge decay and traffic-driven reinforcement.
Spatio-temporal forecasting method using sheaf diffusion to model heterogeneous local interactions in systems, reformulating prediction as information flow over locally structured spaces.
Federated learning approach using personalized submodels aligned with client computational budgets, addressing structural diversity and representational adaptation across heterogeneous devices.
Modular neural architecture for learning three-valued logic end-to-end across mathematical domains without external symbolic solvers.
Memory-Enhanced Dynamic reward Shaping (MEDS) framework for reinforcement learning that reduces failure pattern recurrence in LLM training.
Framework for training deep learning models on non-differentiable scientific metrics using knowledge-guided learning and Minkowski Image Loss.
Research on optimizing reinforcement learning with verifiable rewards for LLMs using low-rank trajectory modeling to reduce computational training overhead.
Dynamic mixture-of-experts architecture for graph neural networks that adapts expert computation per node based on difficulty.
Diffusion language model for goal-directed molecular generation satisfying heterogeneous constraints like protein compatibility and drug properties.
Comparison of KV-cache compression strategies for transformers showing quantization outperforms rank reduction across multiple model sizes.
Analysis of per-sample forgetting dynamics during fine-tuning of image classifiers, examining architecture dependence of retention patterns.
Continuous-time flow models trained with adversarial objectives using learned discriminators instead of fixed MSE criteria.